Data wrangling

Data wrangling

In real-world data analysis, your data will likely:

• Have missing/possibly incorrect values
• Be in a format unsuitable for data analysis
• Be spread across multiple files, possibly of different types
• Need re-shaping or summarization to draw meaningful conclusions

Fortunately, pandas can help you with all of this!

Subsetting

• Frequently, you’ll have way more data than you need!
• Datasets can be reduced in size by indexing and subsetting
• Let’s read in the colleges dataset as a demo

import pandas as pd

full_url = 'http://personal.tcu.edu/kylewalker/data/colleges.csv'
full = pd.read_csv(full_url, encoding = 'latin_1')
full.shape

By column name

• Let’s drop most of the columns in the dataset

cols_to_keep = ['INSTNM', 'STABBR', 'GRAD_DEBT_MDN_SUPP']
debt = full[cols_to_keep]
debt.columns = ['name', 'state', 'debt']
debt.head()

By row position

• Data frames can be sliced like lists and strings

debt[0:10]

By row or column index

• Selecting by row or column index available in the .ix[] method (note the brackets)

ex1 = debt.set_index('name')
ex1.ix['Amridge University':'Alabama State University']

By value

• Often, you’ll want to keep rows that have a certain column value, or exclude rows based on that value
• Statements use conditional operators; can be combined with & (and) and | (or)
• .isin() method: checks to see if value is in list of values

By value

debt1 = debt[debt['debt'] != 'PrivacySuppressed']
debt1.head()

By value

tx_debt = debt[(debt['debt'] != 'PrivacySuppressed') & (debt['state'] == 'TX')]

# Alternatively, use the .query() method

# tx_debt = debt.query('debt != "PrivacySuppressed" & state == "TX" ')

tx_debt.head()

By value

states = ['OK', 'NM', 'TX', 'LA']

sw_debt = debt[(debt['debt'] != 'PrivacySuppressed') & (debt['state'].isin(states))]

sw_debt.head()

Creating new columns

• New columns can be created based on specified values, or as derivatives of other columns, using mathematical operators or the .assign() method
• Let’s demo with a simulated data frame:

import numpy as np
np.random.seed(1983)

df1 = pd.DataFrame({'col1': np.random.randint(1, 100, 10), 
                    'col2': np.random.randint(1, 100, 10), 
                    'col3': np.random.randint(1, 100, 10)})

df1

Creating new columns

# Must use index-based labeling for this
df1['col4'] = df1['col1'] + df1['col2']

df1.head()

df2 = df1.assign(col5 = df1['col3'] / df1['col4'])

df2.head()

dtype conversion

• To do numerical analysis, our numeric data have to be stored as numbers!
• To convert: use the .astype() method
• Note: SettingWithCopyWarning

sw_debt['debtnum'] = sw_debt.debt.astype(float)

sw_debt.head()

Missing data

• Commonly, all of the data you need will not be found in your data set!
  
• Possible solutions:
  • Delete all rows that have missing data
  • Fill in missing data with a specified value
  • Interpolate missing values

Missing data

• .dropna() method: delete all rows (or columns) that have any missing values (NaN in pandas)

sw2 = sw_debt.dropna()

sw2.head()

Missing data

• .fillna() method: fill in missing data with a specified value

sw3 = sw_debt.fillna(sw_debt.mean())

sw3.head()

Group-wise data analysis

• Thus far, we’ve focused on characteristics of data within a particular group
• Common question: how do characteristics vary by group?
  
• In pandas: .groupby() method!

Split-apply-combine

• Wickham (2011): the “split-apply-combine” model of data analysis

Process:

• Data are split by some characteristic into groups
• We apply a function to each of the groups
• The resultant data are combined back into a single dataset

.groupby() in pandas

sw_grouped = sw2.groupby('state')

sw_grouped['debtnum'].mean()

# Result

state
LA    15876.255319
NM    16237.466667
OK    17030.860759
TX    15009.426582

Grouped visualization in seaborn

import seaborn as sns
%matplotlib inline

sns.boxplot(x = 'state', y = 'debtnum', data = sw2, orient = 'v')

Grouped visualization in seaborn

• Faceting or small multiples: breaking down a plot by a grouping variable into multiple plots

g = sns.FacetGrid(data = sw2, col = 'state', col_wrap = 2)
g.map(sns.distplot, 'debtnum')

Grouped visualization in seaborn

• We can also get creative with our code!

for st in sw2.state.unique(): 
    data = sw2[sw2.state == st]
    sns.kdeplot(data.debtnum, shade = True, label = st)

Merging data

• Commonly, you’ll have data in two - or multiple! - datasets that you’ll want to combine into one
• Simulated data:

np.random.seed(123456)

m1 = pd.DataFrame({'type': ['a', 'b', 'c', 'd', 'e', 'f'], 
                  'ind1': np.random.randint(1, 100, 6), 
                  'ind2': np.random.randint(1, 100, 6)})

m2 = pd.DataFrame({'type': ['a', 'b', 'c', 'd', 'e', 'f'], 
                  'ind3': np.random.randint(1, 100, 6), 
                  'ind4': np.random.randint(1, 100, 6)})

The .merge() method in pandas

m3 = m1.merge(m2, on = 'type')

Types of merges in pandas

• Options for merging (the how parameter): 'inner' (default), 'left', 'right', and 'outer'
• Simulated data:

m4 = pd.DataFrame({'type': ['d', 'e', 'f', 'g', 'h', 'i'], 
                  'ind5': np.random.randint(1, 100, 6), 
                  'ind6': np.random.randint(1, 100, 6)})

Inner merges

m5 = m1.merge(m4, on = 'type', how = 'inner')

Left merges

m5 = m1.merge(m4, on = 'type', how = 'left')

Right merges

m5 = m1.merge(m4, on = 'type', how = 'right')

Outer merges

m5 = m1.merge(m4, on = 'type', how = 'outer')

The “shape” of data

• Long (“tidy”) data:
  • Each variable forms a column;
  • Each observation forms a row;
  • Each type of observational unit forms a table
• Wide data: column headers represent values, not variable names

Example: World Bank data

• Long format:

from pandas.io import wb
countries = ['ZA', 'BR', 'US']
urban = wb.download(indicator = 'SP.URB.TOTL.IN.ZS', 
                    country = countries, start = 2007, 
                    end = 2011).reset_index()
urban.head()

Long to wide

• .pivot() method in pandas

urban_wide = urban.pivot(index = 'year', columns = 'country', values = 'pcturban')
urban_wide.head()

Plotting “wide” data

urban_wide.plot()

Wide to long

• pd.melt() function in pandas

urban_long = pd.melt(urban_wide.reset_index(), id_vars = 'year', 
                     var_name = 'country', value_name = 'pcturban')

urban_long.head()